l'RONTlSPIECE. Unmagnified high-altitude image ot tlremerton, wasmngwn \scale 1: 135,000). Original image recorded on false-color- reversal film.

CHARLES E. OGROSKY University ofWashington Seattle, WA 98195 Population Estimates from

A high degree of association between urban area population and four variables representing urban size and relative dominance, as measured on high-altitude satellite imagery, indicated that such imagery may be useful for estimating urban population.

(Abstract on next page)

INTRODUCTION completely survey the population of the United States and its characteristics. Such UMERous examples of the useof conven­ methods of data collection have been criti­ N tional aerial photographs in analyzing cized as being overly costly and time con­ urban problems have been reported. The suming, given the accuracy of the resultant purpose of this study was to investigate the data. Use ofaerial imagery for collecting cer­ suitability of high-altitude satellite imagery tain types of census information has been for estimation of regional urban population. suggested as a procedure which can success­ fully supplement existing techniques.! Al­ CENSUS METHODS though much of the data presently collected Both field and mail questionnaire enumer­ by traditional methods would not be directly ation methods are frequently employed to obtainable from remotely sensed images, 707 708 PHOTOGRAMMETRIC ENGINEERING & , 1975

ABSTRACT: The suitability ofmedium ground resolution, high altitude satellite imagery as a data source for intercensal population esti­ mates was evaluated. Four variables representing urban size and relative dominance were measured directly from unmagnified high­ altitude, color-infrared transparencies for each of18 urban test sites in the Puget Sound region. Linear relationships between each vari­ able and the population of the test sites were established, the strongest association being between the logarithm ofthe area ofthe site (Y), r 2 = 0.964. Use ofallfour independent variables in a multiple linear regression equation resulted in a slightincrease in the r2 value, to 0.973. The high degree ofassociation indicated that relatively low ground resolution images may be useful for certain intercensal es­ timating purposes. Further research is needed to define additional variables which may be useful for predictive purposes.

some information is immediately obvious or L i - number ofsurface links from test may be inferred from known relationships to site i to other urban plaes was obtained patterns of physical phenomena. One of the directly from the large-scale imagery. best-established uses of aerial photographs Pj - population ofthe nearest larger urban area in urban area management is for collection of j was recorded from census records. Dij - highway distance from test site i to nearest information about population dispersal. Ad­ larger urban areaj measured on highway vantages of such uses are- maps. • Intercensal Estimates. Changes in urban Ai - urbanized area of test site i obtained from population patterns which occur within the imagery. existing census cycle can be detected. • Wide application. Although expensive, un­ The variables were combined using a step­ iform aerial coverage of large regions has wise linear regression procedure to yield a potential applications in numerous discip­ regression equation of the form: lines. • Permanent Record. Remote sensing pro­ vides a relatively nonselective, permanent recording of physical phenomena and, by inference, of the interrelationships bet­ where k is a constant and b 1..4 are the coeffi­ ween those phenomenaandtheir total envi­ cients of each independent variable. Coeffi­ ronment. • Ease ofAccess. Aerial survey is well suited cients of determination (r2Y of 0.90 and 0.77 to coverage of remote areas and for data were obtained for each year studied (1953 collection where political, social or and 1963). Analysis showedthatAi correlated economic conditions make field enumera­ particularly well with Pi for smaller urban tion impractical. places but that other variables were needed to more accurately estimate populations of EXISTING RESEARCH larger urban areas. Most investigations of aerial imagery as a The results obtained by Holtzet al.,3 HSU,2 data source for studies of urban phenomena Lindgren and others4 indicate that data col­ have used large-scale, conventional black­ lection from conventional aerial images may and-white or color photographs. HSU,2 for ex­ be a viable alternative to ground-based tech­ ample, employed images with an average niques for some estimating purposes. As scale of about 1:5000 for derivation of inter­ WobberS suggests, the increasing availability censal population estimates of urban areas of high-altitude and satellite imagery for near Atlanta, Georgia. Holtz et al,3 using aer­ non-military purposes makes evaluation of ial photographs, attempted to develop a such images as potential data sources appro­ model for predicting existing and future priate. Use of small-scale satellite imagery, populations for urban places of2500 or more such as from the Resources Technol­ persons in the Tennessee River Valley. Five ogy Satellite (ERTS) or from SKYLAB, may variables were recorded for each test site. be desirable for two reasons: it provides a The dependent variable, urban population synoptic view ofregional patterns and trends (Pi) ofeach test site i, was recorded from cen­ and repetitive coverage which has proven sus records. The four independent variables useful in analysis of dynamic natural and were recorded from a variety of sources: man-made processes. POPULATION ESTIMATES FROM SATELLITE IMAGERY 709

RESEARCH OBJECTIVES undeveloped military reservations in urban areas. The outlined area was measured with Imagery obtained from very high-flying a polar planimeterand the data converted to such as the U-2 and from satellite sen­ square miles. sor platforms incorporates, in effect, a spatial • Surface Links. (LJ Rather than simply re­ filter due to sensor system resolution limits. cording the total number of highway and The effect of such filtering is to eliminate rail links leading from each urban test site, spatially high frequency (small) phenomena differentiation by link significance was in­ while passing lower frequency (large) ob­ corporated byassigning arbitrary values of4 jects. Although larger-scale, high-ground­ to multiple lane highways, 2 to secondary resolution images may be available, use ofspa­ roads and 1 to railroads. The information was obtained directly from the images. tially 'filtered' records may be beneficial. • Urban Area of Nearest Larger Neighbor­ Such use requires concentration on synoptic, hood (NLN). (Aj) To restrict data collection or neighborhood-regional, elements and may to aerial images, the population of the enhance gross urban structure which might nearest larger neighborhood variable (Pj ) otherwise be masked by a wealth of high­ used by Holtz et aI.3 was changed to area. frequency detail. Thus, on most available The area in square miles was recorded as for small-scale civilian imagery, individual 'Urban Area' above. However, high altitude structures are not readily discernible and imagery for Vancouver, British Columbia conventional approaches to urban population and for San Francisco, nearest larger neigh­ boring urban areas for Bellingham and Seat­ estimation, such as individual house counts tle, was not available and other documents or structural density evaluation, are inap­ were used to obtain the area ofthese cities. propriate. However, residential, industrial • Highway Distance to Nearest Larger and commercial areas are visible in gross Neighbor. (Dij) The distance, in miles, from form. This study is an effort to determine each test site to its nearest larger urban whether the physical phenomena recorded neighbor was measured directly from the in relatively low ground-resolution images imagery. Ferry routes were used when ap­ can be successfully related to variations in propriate. urban population. DATA ANALYSIS DATA ACQUISITION The four independent variables (X) were Recent census data6 was reviewed and 18 first plotted singly against the dependent incorporated urban places in the Puget Sound Basin (Figure 1) were selected for analysis. The test sites represent the most intensely populated portions of the region, including suburban communities on the urban-rural fringe. High-altitude color­ infrared transparencies (9 by 9 inches) at a scale of 1:135,000 (Frontispiece) were the primary data source.7 To simulate use of N medium-ground resolution satellite imagery, t such as from the SKYLAB missions, test sites were limited to urban places of 10,000 or more persons and no magnification of the high-altitude images was used during data collection. Port Angeles The four independent variables measured by Holtz et al. 3 were retained for this study, but were modified to make them more consis­ tent with the theories of urban settlement dispersal and to make it possible to obtain all dafa directly from the imagery rather than from a variety of sources. The data, derived from the transparencies, are listed in Table 1. • Kent • Auburn TACOMA • Urban Area. (Ai) Gross urban form was re­ • corded for each test site by tracing the Puyallup boundaries of residential, commercial and a 25mi. industrial areas on an overlay. Allowance was made for major parks, water bodies and FIG. 1. Puget Sound Region urban test sites. 710 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING, 1975

TABLE 1. YDATA FROM BUREAU OF CENSUS REPORTS. XDATA COLLECTED FROM HIGH-ALTITUDE AERIAL IMAGERY

N.L.N. Distance to Population Area Area N.L.N. Test Site (lOOO's) (mi 2) Links (mi 2) (mi 2) Seattle 531 95.11 44 109.27 827 Tacoma 155 26.92 24 95.11 31 Bellevue 61 13.71 19 95.11 11 Everett 54 11.71 14 95.11 29 Bellingham 39 10.12 16 79.05 53 Bremerton 35 6.40 11 95.11 15 Renton 25 5.76 15 95.11 13 Edmonds 24 4.13 10 95.11 14 Olympia 23 5.17 15 26.92 28 Auburn 22 5.03 12 26.92 10 Kent 21 4.13 13 5.93 6 Mercer 1. City 19 3.99 6 13.71 7 Lynnwood 17 3.63 5 4.13 4 Mountlake Terrace 17 3.54 8 4.13 5 Port Angeles 16 3.13 7 6.40 76 Kirkland 15 1.95 6 13.71 4 Puyallup 15 2.36 8 36.92 9 Redmond 11 2.04 7 1.95 4 population data (Y); the results are portrayed A positive association exists between all of graphically in Figures 2 through 5. Linear the independent data categories and log Pi' relationships of varying strengths between Although the coefficient of determination all independent variables and urban popula­ (0.973) is slightly higher than that obtained tion (P) are shown by regression lines and the by Holtz et al. 3 for either year of their study coefficients of correlation (r) are indicated. (0.902 and 0.774), the same general relations

Three ofthe variables (Pj , Ai and Dij) showed prevailed. By far the best single estimator of the best linear relationships when logarith­ urban population which was visible on high mic transformations were employed. The altitude imagery was log Ai' Because of their correlation matrix (Table 2) shows the high non-orthogonality, addition of the other in- degree of association between several of the variables on the basis ofsimple linear regres­ sion. Although log Ai and Li each correlated highly with log Pi (r2 of 0.964 and 0.897), . 500 there were high degrees of association be­ tween independent variables as well. For ex­ 2 R = .982 ample, log Ai and L i showed an r value of 0.874. , _100 The dependent variable, log Pi' and all four !S- independent variables were analyzed using c ~ 50 the Biomedical Computer Programs o '3 (BlOMED) series multiple linear regression g- program 02R. The stepwise procedure adds o.. independe~tvariables to the regression equa­ .s tion until all data classes with an F value of 1 5 10 0() 100 more than 0.01 have been used. In this study, FIG. 2. Area (Log Ai) plotted against all four variables were added, yielding the population (Log Pi)' regression equation: TABLE 2. CORRELATION MATRIX log Pi = 0.6896log Ai + 0.0106L + 0.0004A i j Variable Log Pi Log Ai Li Aj Log Dij + 0.0143log D + 0.7724 ij Log Pi 1.000 log Ai 0.982 1.000 where 0.7724 is a constant analagous to 'a' (Y Li 0.948 0.935 1.000 intercept) in the two-dimensional linear­ Aj 0.739 0.742 0.651 1.000 log Dij regression equation Y = a + bX. 0.729 0.713 0.743 0.540 1.000 POPULATION ESTIMATES FROM SATELLITE IMAGERY 711

R = .729 500

~_IOO R = .948 ~ ~ -... I 100 ! ~- .3 10 l--"-'-;:"'~ __~~_~~~_- 1 10 50 100 c 50 """ 1000 .~ Highway Distance to Nearest larger Neighbor (Log Dq) - Miles 15 :; 0. FIG. 5. Highway distance to nearest 0 ... l>. . . I larger neighbor (Log dij) plotted against 8' population (Log Pi). -' 10 12 24 36 48 Surface Transport Links (~) - Highway and Railroad jective definition and measurement of inde­ FIG. 3. Surface transport links (Lil plotted pendent variables, investigators may be able against population (Log PJ to obtain reasonably accurate population es­ timates for politically, physically or econom­ ically inaccessible areas at reasonable cost. """ Group 1 Assuming that wider availability of high­ 1 R '" .739 quality civilian satellite imagery is probable, J Group 1 further work should be done on defining ;:- 100 other variables which may show even .~ : stronger associations with urban structure .. '" Group 3 and from which other relationships may be J Group .( inferred. In particular, relationships of the .3 10 0 100 nature discovered in this and similar studies Area of Nearest lorger'"Neighbor (Ai) - Sql)Clre Miles should be compared to define spatial (re­ This figure suggests a regional hierorchy of vrban places gional)differences. Temporal differences may based on the area of the largest neighboring urbon settlement and population. be revealed by analysis ofcurrent and histor­ ical aerial coverage for a variety of test re­

Extl'(lregionol Group 1 Group 2 Group 3 Group 4 gions. In this manner, investigators may be Vancouver Bellingham Kirkland able to develop enumeration techniques EV'IIrelt ... /.v.ercer I. City BelleYue ~Redmond which may lead to increasingly powerful Son FronchcQ- Seattle Bremerton_Port Angeles Ren'Ot'l lynnwood theoretical explanations about the arrange­ ~ ~ Edmor'lds Mountlake Tel'T(lC;:e ment and growth of urbanized areas. ~ Aubv," To<:cmo Kent ~ Puyallup Olympia TABLE 3. INCREASE IN COEFFICIENT OF FIG. 4. Area of nearest larger neighbor DETERMINATION (R2) WITH ADDITION OF (Aj) .plotted against population (Log Pi)' INDEPENDENT VARIABLES Step X Variable Increase in Number Entered R2 R2 dependent variables, which were based on 1 logA; 0.9647 0.9647 more theoretical constructs about the hierar­ 2 L; 0.9716 0.0068 chical dispersion and connection of urban 3 Aj 0.9725 0.0010 places, resulted in an increase of r 2 of less 4 log Dij 0.9727 0.0002 than 0.01 (Table 3).

CONCLUSIONS ACKNOWLEDGMENT The high degree of linear association be­ The assistance of Mr. Frank Westerlund of tween Puget Sound region population and I the Remote Sensing Applications Labora­ four independent variables, most notably tory, Dept. of Urban Planning, Univ. of urban area, measured directly from unmag­ Washington, Seattle in obtaining the high­ nified high-altitude imagery (scale altitude imagery is gratefully acknowledged. 1: 135,000) indicated that satellite imagery of approximately the ground resolution ob­ REFERENCES tained on SKYLAB missions could be a useful 1. For example, B. S. Wellar, "Remote Sensing indicator of regional urban structure. By ob- and Urban Information Systems," 712 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING, 1975

Photogrammetric Engineering, 39 (10), L. Mumbower and J. Donoghue, "Urban Pov­ 1041-1050. erty Study," Photogrammetric Engineering, 2. S. Y. Hsu, "Population Estimation," Photo­ 33 (6), 610-618. grammetric Engineering, 37 (5), 449-454. D. M. Richter, "Sequential Urban Change," 449-454. Photogrammetric Engineering, 35 (8), 3. R. K. Holtz, D. L. Huff and R. C. Mayfield, 764-770. "Urban Spatial Structure Based on Remote 5. F. J. Wobber, "Orbital Photos Applied to the Sensing Imagery," in R. K. Holtz, Ed., The Environment," Photogrammetric Engineer­ Surveillant Science, (Boston: Houghton Mif­ ing, 36 (8), 852-864. flin, 1973), 375-380. 6. U. S. Bureau ofthe Census, Census ofPopula­ 4. D. T. Lindgren, "Dwelling Unit Estimation tion, 1970. Vol. 1, Characteristics ofthe Popu­ with Color IR Photos," Photogrammetric En­ lation, Part 49, Washington. (Washington, gineering, 37 (4), 373-377. D.C.: G.P.O., 1973).

Symposium on Close-Range Photogrammetric Systems

Champaign-Urbana, Illinois July 28 - August 1, 1975

The Symposium on Close-Range Photogrammetric Systems, sponsored by the Close-Range Committee and the Computational Photogrammetry Committee of the American Society ofPhotogrammetry in cooperation with ISP Commission V and the Univer­ sity ofIllinois at Urbana-Champaign, will be held at the Ramada Inn Convention Center, 1501 South Neil Street, Champaign, Illinois 61820.

Session topics will include:

Architectural Applications Engineering and Industrial Applications Biostereometrics Underwater Mapping Systems & Applications Holographic and Microscopic Systems & Applications Computational Procedures and Software (2 sessions) Photogrammetric Equipment Systems Manufacturers-Users Forum

A number ofphotogrammetric equipment manufacturers, government agencies (including the Naval Photographic Center) and universities will exhibit the latest equipment systems and other items pertaining to close-range photogrammetry. Registration fees will be $30 for ASP members and $35 for non-members. For information on the Symposium contact Dr. H. M. Karara, 3129 Civil Engineering Building, University of Illinois, Urbana, Illinois 61801. Tel: (217) 333-4311.